Advancing towards semi-automatic labeling of GPR images to improve visualizations of pipes and leaks in water distribution networks using multi-agent systems and machine learning techniques

Abstract

Critical infrastructures such as water distribution networks (WDNs) require reliable and affordable information at a reasonable cost to address challenges that can negatively affect their operation. Inadequate knowledge about WDN assets and their state of health presents challenges for essential activities such as network modeling, operation, assessment, and maintenance. This work seeks to increase the availability of WDN asset data through improved interpretability of GPR images. The semi-automatic labeling approach presented here expands upon existing multi-agent image-cleaning methods and feature characterization techniques. The division of a pre-processed image, in the form of a matrix, into a grid of smaller blocks allowed the identification of relevant features using density of nonzero values in the blocks; this approach, conducted manually in this proof of concept, can provide a basis for training an intelligent system (e.g., a convolutional neural network) to extract the families of interest and eliminate noise. Thus, this research expands this methodology to advance towards automatic detection of pipes and leaks and easily visualize the data. In this paper, 3D visualizations of WDN assets have been created to demonstrate the usefulness of this semi-automatic process in delivering easily-interpretable GPR data for managers and operators of WDNs.